1. technology and computing
2. internet technology
3. web search

Why Searchers Switch: Understanding and
Predicting Engine Switching Rationales
Qi Guo1, Ryen W. White2, Yunqiao Zhang2, Blake Anderson2, and Susan T. Dumais2
1
Mathematics and Computer Science, Emory University, Atlanta, GA 30322, USA
2
Microsoft Corporation, One Microsoft Way, Redmond, WA 98052, USA
qguo3@emory.edu, {ryenw,yuzhan,blakean,sdumais}@microsoft.com
ABSTRACT
Search engine switching is the voluntary transition between Web
search engines. Engine switching can occur for a number of reasons, including user dissatisfaction with search results, a desire for
broader topic coverage or verification, user preferences, or even
unintentionally. An improved understanding of switching rationales allows search providers to tailor the search experience according to the different causes. In this paper we study the reasons
behind search engine switching within a session. We address the
challenge of identifying switching rationales by designing and
implementing client-side instrumentation to acquire in-situ feedbacks from users. Using this feedback, we investigate in detail the
reasons that users switch engines within a session. We also study
the relationship between implicit behavioral signals and the
switching causes, and develop and evaluate models to predict the
reasons for switching. In addition, we collect editorial judgments
of switching rationales by third-party judges and show that we can
recover switching causes a posteriori. Our findings provide valuable insights into why users switch search engines in a session and
demonstrate the relationship between search behavior and switching motivations. The findings also reveal sufficient behavioral
consistency to afford accurate prediction of switching rationale,
which can be used to dynamically adapt the search experience and
derive more accurate competitive metrics.
Categories and Subject Descriptors
H.3.3 [Information Storage and Retrieval]: Information Search
and Retrieval–selection process, search process.
General Terms
Algorithms, Measurement, Experimentation, Human Factors.
Keywords
Search engine switching, predicting switching rationales.
1. INTRODUCTION
Search engines facilitate rapid access to information on the Web.
A user’s decision regarding which engine to use most frequently
(their primary engine) can be based on factors including reputation, familiarity, effectiveness, interface usability, and satisfaction
[18], and can significantly impact their levels of search success
[22][24]. Similar factors can influence a user’s decision to switch
engines, either for a particular query if they are dissatisfied with
results or seek broader topic coverage, or for specific task types if
another engine specializes in such tasks, or more permanently due
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise,
or republish, to post on servers or to redistribute to lists, requires prior
specific permission and/or a fee.
SIGIR 2011, July 24–28, 2011, Beijing, China.
Copyright 2011 ACM 1-58113-000-0/00/0010…$10.00.
to unsatisfactory experiences or relevance changes [23]. The barrier to switching Web search engines is low and multiple engine
usage is common. Indeed, prior work in this area suggests that
70% of Web searchers use multiple search engines [24].
Previous work on switching has examined promoting multiple
search engine use [24], characterizing both short- and long-term
engine switching behavior [22][23], predicting when users will
switch [12][15][22], developing metrics for competitive analysis
of engines using estimated user preference and user engagement
[13], or building conceptual and economic models of search engine choice [17][21]. Despite the economic significance of engine
switching to search providers, and its prevalence among search
engine users, little is known about users’ rationales for switching,
the search behaviors that may relate to different rationales, or the
features that are most useful in automatically predicting different
switching causes. This is in part due to difficulty in reliably identifying switching rationales.
Improved understanding of switching rationales is important and
has several applications. If search engines could predict that a
searcher was about to switch along with the reason for that switch,
they could adapt the search experience accordingly. For example,
if a user searching for [butterfly flight patterns] is about to switch
because they could not find what they were looking for (i.e., the
reason is dissatisfaction), then the search engine can intervene and
provide help such as additional query support. Conversely, if a
user searching for [bellevue apartments] is about to switch to
check for additional offerings (i.e., the reason is topic coverage or
verification), an intervention could be annoying. Even in offline
settings, search providers can improve the search experience
through a better understanding of switching rationales. For example, queries resulting in dissatisfaction-related switches can be
identified and analyzed to improve search quality.
In this paper, we focus on understanding users’ motivations for
search engine switching within a single search session. We address the challenge of identifying switching motivations by implementing and deploying a browser extension to capture search
interaction and acquire explanations from searchers in situ (i.e., at
the time they switch search engines). The in-situ explanations
provide first-hand insights about the reasons for switching, and
can be used to understand the relationships between patterns of
search interaction and the switching causes. We investigate the
effect of different interaction features (derived from the query,
pre-switch user behavior, and post-switch behavior) on the accuracy of models to automatically predict the reason for a switch.
The main contributions of this paper include:
•
The implementation and deployment of client-side instrumentation to collect in-situ searcher explanations that enables
the study of session-level switching rationales.
•
An in-depth analysis of session-level engine switching rationales and their correlations with behavioral signals.
•
An investigation and evaluation of predicting switching
causes with various session-level behavioral features.
This study is the first research to understand switching causes insitu and to predict the reasons for search engine switches, laying
the groundwork for more extensive future work on engine switching and search satisfaction.
The remainder of this paper is structured as follows. Section 2
outlines previous work on predicting query difficulty, user satisfaction and frustration, and characterizing search engine switching
behavior. Section 3 provides an overview of the study. Section 4
describes the design and implementation of our client-side instrumentation for collecting in-situ explanations. In Section 5 we
characterize the session-level switching causes, and their correlations with search behavior. In Section 6, we investigate prediction
of switching rationales, varying features used. In Section 7, we
explore using editorial judgments to approximate the in-situ explanations. We discuss our findings and their implications in Section 8 and conclude in Section 9.
2. RELATED WORK
Three lines of work are most relevant to our research: (i) predicting query performance, (ii) user satisfaction and frustration, and
(iii) characterizing and predicting engine switching behavior.
Research on predicting query performance has been conducted to
understand differences in the quality of search results for different
queries. Such predictions can be used to devote additional resources or use alternative methods to improve search results for
difficult queries. While it has been shown that using different
query representations [4] or retrieval models [3] improves search
performance, it is more challenging to accurately predict which
methods to use for a particular query.
Measures such as query clarity [7], Jensen-Shannon divergence
[6], and weighted information gain [25] have been developed to
predict performance on a query (as measured by average precision, for example). Guo et al. [10] used interaction features, including switching features, to predict query performance. Leskovec et al. [16] used graphical properties of the link structure of the
result set to predict the quality of the result set and the likelihood
of query reformulation. Teevan et al. [20] developed methods to
predict which queries could most benefit from personalization.
Hassan et al. [11] developed methods to predict search success on
a session-level. Feild et al. [8] developed methods to predict user
frustration, and showed that features capable of accurately predicting switching events were also highly predictive of frustration.
Some research has examined search engine switching behavior.
Early research by Mukhopadhyay et al. [17] and Telang et al. [21]
has used economic models of choice to understand whether people developed brand loyalty to a particular search engine, and how
search engine performance (as measured by within-session
switching) affected user choice. They found that dissatisfaction
with search engine results had both short-term and long-term effects on search engine choice. Juan and Cheng [13] described
some more recent research in which they summarize user share,
user engagement and user preferences using click data from an
Internet service provider. They identify three user classes (loyalists to each of the two search engines studied and switchers), and
look at the consistency of engine usage patterns over time.
Heath and White [12] and Laxman et al. [15] developed models
for predicting switching behavior within search sessions using
sequences of user actions (e.g., query, result click, non-result
click, switch) and characteristics of the pages visited (type of page
and dwell time) as the input features. Heath and White [12] used a
simple threshold-based approach to predict a switch action if the
ratio of positive to negative examples exceeded a threshold. Using
this approach they achieved high precision for low recall levels,
but precision dropped off quickly at higher levels of recall. Working with the same data, Laxman et al. [15] developed a generative
model based on mixtures of episode-generating hidden Markov
models and achieved much higher predicative accuracy.
White et al. [24] developed methods for predicting which search
engine would produce the best results for a query. For each query
they represented features of the query, the title, snippets and
URLs of top-ranked documents, and the results set, for results
from multiple search engines, and learned a model that predicted
which engine produced the best results for each query. The model
was learned using a large number of queries for which explicit
relevance judgments were available. One way in which such results can be leveraged is to promote the use of multiple search
engines on a query-by-query basis, using the predictions of the
quality of results from multiple engines. White and Dumais [22]
characterized search engine switching through a large-scale survey and built predictive models of switching based on features of
the pre-switch query, session, and user. White et al. [23] modeled
long-term engine usage over a six-month period, and identified
three user classes: (i) those who do not switch, (ii) those who
switch at some time, and (iii) those who switch back and forth
between different search engines.
We extend previous work on search engine switching by focusing
on understanding the rationales for search engine switching in a
single search session in depth, and on predicting switching rationales given features of the query and pre-/post-switch behavior.
3. STUDY OVERVIEW
We begin by providing definitions used throughout the paper. We
then provide an overview of the reasons for engine switching that
were identified in previous work, and present the research questions that we address in this paper.
3.1 Definitions
Some terms are formally defined as below.
DEFINITION 1. A search session is a sequence of user activities
that begins with a query, includes subsequent queries and URL
visits, and ends with a period of inactivity. URL visits include
both clicks on the search engine result page (SERP clicks) and
post-SERP navigation. We used 30 minutes of user inactivity to
mark session boundaries; similar timeouts have been used previously to demarcate search sessions in Web log analysis [22].
DEFINITION 2. A search engine switching event is a pair of consecutive queries that are issued on different search engines within
a single search session. In our definition of a switching event,
navigational queries for search engine names (e.g., search on Yahoo! for [google], [google.com], etc.) are regarded as part of the
act of switching and not as the pre- or post-switch query. For example, if a Yahoo! user searches for [snowshoes], then for
[google] and switches to Google, and then searches for [snowshoe
merchants], then the pre-switch query is [snowshoes] (and not
[google]), and the post-switch query is [snowshoe merchants].
DEFINITION 3. A search goal is an atomic information need,
resulting in one or more related search queries issued to accomplish a single discrete task [11]
3.2 Reasons for Search Engine Switching
In earlier research on characterizing and predicting search engine
switching behavior, White and Dumais [22] surveyed 488 users
regarding their experiences with search engine switching. They
asked respondents to provide retrospective rationales for switching by selecting at least one explanation from a list of possible
reasons. For reader reference, in Figure 1 we present the response
breakdown reported in their study.
As we can see from the figure, there are three classes of reasons:
dissatisfaction (DSAT) with the results in the original engine (dissatisfaction, frustration, expected better results; shown in red in
Figure 1 and totaling 57%), verifying or finding additional information (coverage/verification, curiosity; in green and referred to
as “Coverage”, totaling 26%), and user preferences (destination
preferred, destination typically better; in blue, totaling 12%).
Destination
typically better
(9%)
Unintentional
(2%)
Other (3%)
Dissatisfaction
(24%)
User Preferences
demonstrated feasibility in asking human judges to label search
success [11]. However, switching rationales are more subjective
and personal since they relate not only to searcher goals and result
relevance, but also to other factors such as long-term searcher
preferences and browser settings. As a result, it is challenging to
identify real switching rationales by examining log data alone.
Another option is to ask searchers why they switch search engines. White and Dumais [22] asked people to summarize their
reasons for switching using a retrospective questionnaire. While
this provides some interesting insights, retrospective surveys do
not always align with actual behavior and the corresponding behavioral data is not available. Thus, we chose to ask searchers
about their switching behavior in situ when they switch engines.
To obtain these in-situ switching assessments we implemented
and deployed a browser add-on, called SwitchWatch, which presents a short questionnaire to the user at the time of a switch between Google, Yahoo!, or Bing. This questionnaire elicits switching rationales directly from the user at switch time. Figure 2
shows the questionnaire which contains questions about the search
task and switching rationale (as described in more detail below).
Destination
preferred (3%)
DSAT
Curiosity (17%)
Coverage
Coverage or
verification
(9%)
Expected
better results
(23%)
Frustration
(10%)
Figure 1. Reasons for search engine switching.
In this study, we collected data about switching rationales at the
time the switches happen and develop models to distinguish between DSAT- and coverage-related switches using session-level
evidence. We focus on these classes since they are most common,
are directly actionable by search providers, and unlike preferencerelated switching, do not require knowledge of users’ long term
behavior, which may not be available to the engine at query time.
3.3 Research Questions
Our study answers a number of research questions:
1. Why do searchers switch search engines?
2. Which behavioral signals are associated with different causes?
3. How accurately can we predict the causes of engine switching?
Answers to these questions can help us better understand switching and help search providers improve the user experience in accordance with searchers’ motivations for switching or derive more
sophisticated competitive metrics for comparing search engines.
4. IDENTIFYING RATIONALES
The first challenge we addressed was collecting switching rationales and their associated search behaviors. One possibility is to
design difficult search tasks and ask participants to try to complete
these tasks (e.g., [1][8]). For this approach, the challenge lies in
the design of tasks that will induce switching behavior naturally.
Since switching behavior is rare [22], it is difficult to collect a
sufficient amount of switching data in a laboratory setting if participants are not instructed to switch but this will bias the data.
Another possibility is to sample from search engine logs and recruit human judges to label sessions. Previous research has
Figure 2. Example SwitchWatch questionnaire, where the
pre- and post-switch queries are [windows phone 7].
This approach allows us capture the subjectivity or variation in
switching rationales among different users. Not surprisingly, the
main challenge of this method is the slow pace with which labels
are obtained since switching occurs infrequently [22]. The pop-up
survey may also be irritating, and lead to users ignoring the survey, or terminating participation. However, since switching does
not occur often, this is not a big concern. Further, we allow people
to ignore the questionnaire (by clicking on the Ignore button).
We deployed SwitchWatch within Microsoft Corporation for one
month to collect sufficient data. We now provide details on its
implementation and deployment.
4.1 SwitchWatch Implementation
SwitchWatch was implemented as an add-on for the Internet Explorer browser. Once installed, it started automatically every time
a new browser tab was opened and recorded browser activity in
that tab to a remote server. For each visited Web page, we record-
ed the URL, the timestamp, a unique browser window identifier,
and an anonymous user identifier. A separate instance of SwitchWatch ran for each active browser tab and tab focus/blur events
were recorded, allowing us to accurately identify multi-tab usage
and Web page dwell times. Intranet and secure (https) URL visits
were excluded to help maintain user privacy.
Once an engine switch between two of Google, Yahoo!, and Bing
was detected (per definition 2), SwitchWatch displayed a questionnaire in the center of the screen occluding part of the active
browser window. The questionnaire is shown in Figure 2 for a
switch between Google and Bing on query [windows phone 7].
As we see in Figure 2, there are two questions shown to users: (i)
whether the user changed their search goal (to help determine the
extent to which switching was related to goal shifts), and (ii) the
cause of the switch. For the first question, the user can pick one
response from “exactly the same,” “related,” and “different” via
the radio buttons. For the second question, the user can select
multiple reasons that apply. The options include dissatisfaction,
coverage, and unintentional. In addition, two options about user
habit and preferences are included, namely, “the post-switch engine is better for this type of query” and “the post-switch engine is
what I usually use.” These response options were selected based
on the survey responses in [22], and summarized in Figure 1.An
Ignore button is also provided to allow users to skip the pop-up if
they do not want to interrupt their current task to answer.
4.2 SwitchWatch Deployment
We distributed an invitation to deploy the SwitchWatch add-in via
email to approximately 2,200 employees within Microsoft Corporation, including colleagues in affiliated groups, interns and 1,000
randomly-selected full-time employees from across the organization. Invitations were sent to employees with a diverse range of
occupations, from software engineers and testers, to program
managers, paralegals, and administrative staff. 216 employees
participated in the study by installing the add-on on their machine,
for a response rate of around 10%. Privacy restrictions prevented
us from determining the identities or occupations of participants
who accepted the invitation to participate. We ran this study for
approximately four weeks. In each week, we randomly selected a
participant with SwitchWatch installed for the week and rewarded
them with a 50 USD gift card. There were no other usage requirements to be considered for the prize drawing. Our goal was
to retain participants in our study, and make sure that they behaved normally, without forcing them to switch search engines as
they may not do normally in their daily search routine.
5. CHARACTERISTICS OF SWITCHING
We now discuss findings on switching characteristics. First, we
present the overall summary of search engine switching in the insitu logs. Second, we discuss how frequently users change search
queries and search goals when they switch engines, and the relationship between changing search queries and different switch
causes. Finally, we characterize the different switch causes and
their associated search behaviors in the session.
5.1 Definitions
We start by providing some definitions used in this section:
• Same Query (SQ): Identical pre- and post-switch queries.
• Related Queries (RQ): Pre- query and post-switch query share
at least one query term that is not a stop word, but are not SQ.
• Different Queries (DQ): Pre-switch query and post-query do
not share any (non-stop-word) terms.
•
•
•
•
Same Goal (SG): User has the same search goal.
Related Goals (RG): Search goals related but not same.
Different Goals (DG): Search goals are totally different.
Ignored: User dismissed the SwitchWatch questionnaire without providing feedback by clicking Ignore button.
5.2 Data Overview
We now provide more details on the data that were gathered as
part of our experiment. We begin by describing relevant features
of the log data gathered by the SwitchWatch add-in.
In the in-situ log, 20,554 queries were issued on Google, Yahoo!
or Bing by our participants in the four-week duration of the study.
Among all the queries, we observed 1029 switches. We excluded
25 switches that were suggestive of users testing SwitchWatch
(e.g., assessments with queries [test], [hello world], etc.). As a
result, we considered 1,004 instances of search engine switching
events, of which 562 (56%) received in-situ assessments. In the
remaining 45% of switches, participants clicked Ignore, indicating
that they did not wish to offer a reason for the switch at that time.
The 1,004 session-level switches in our set comprise only 4.2% of
all queries, while 107 (49.5%) of the 216 users who installed the
add-in switched search engine within a session at least once (similar to switching rates reported in [22] which was for a much larger
and more heterogeneous sample of search engine users).
5.3 Switch Causes and Goal Changes
In this section, we investigate how the changes of search queries
and search goals relate to the underlying switching rationales. The
inspiration of this analysis came from the observed high percentage of query changes in engine switching events.
5.3.1 Query Changes
We begin by analyzing the breakdown of query changes during
engine switching. As described earlier, query change measured by
the overlap between pre- and post-switch queries. In our logs,
only around 32% of the engine switching events observed had an
identical pre- and post-switch query (SQ), and approximately 50%
of the query pairs shared at least one query term (SQ+RQ), the
remaining 50% of query pairs comprised different queries (DQ).
This raises a question regarding search goal inconsistency during
engine switching, something that we explore next.
5.3.2 Goal Changes
We now examine the relationship between changes of search queries and changes of search goals (i.e., characterized by SG, RG,
and DG as defined earlier). Table 1 shows the breakdown of userreported goal changes with respect to query changes. The table
shows the percentage of switches where participants clicked
Ignore and the fraction of remaining judged (non-ignored) queries
that received each goal change label.
Almost all (98%) of the switches with the same pre- and postswitch query (SQ) share the same search goal (SG). In contrast,
only 20% switches with related pre- and post-switch queries (RQ)
are reported as having related goals (RG), and only 65% of
switches with different pre- and post-switch queries (DQ) are
reported as having different search goals (DG). This difference is
mainly due to the contributions of RQ (77% SG) and DQ (23%
SG) to switches with same goal. This seems reasonable since
users might change the query terms slightly or perhaps make
typos when switching with the same search goal. Nevertheless,
SQ is the best proxy of SG, while RQ could also be a proxy for SG
to increase coverage, and DQ could be a reasonable proxy of DG.
Table 1. Breakdown of goal and query changes.
Ignored
All
SQ
RQ
DQ
SG
RG
DG
45%
65%
9%
25%
27%
39%
60%
98%
77%
23%
1%
20%
12%
1%
3%
65%
There are differences in how often the SwitchWatch dialog is
ignored (by selecting the Ignore button) for different query
changes. The more related the queries, the more likely users are to
provide feedback. The ignore rate of SQ switches is 27%, while
the ignore rate of DQ increases to 60%. Note that we cannot
compute the ignore rates for goal changes (SG, RG, DG) since we
do not have the search goal information for ignored switches
because it was captured explicitly from users by SwitchWatch.
90%
Percentage of switches
Query
change
100%
Goal change
[% judged (non-ignored) queries]
80%
70%
60%
50%
40%
30%
20%
10%
0%
%SQ
%RQ
%DQ
%All
Dissatisfaction
50
39
5
26
Coverage
20
9
0
8
5.3.3 Impact on Switch Causes
Other
9
4
18
13
Now, we study how the change of search queries (as a proxy of
search goal change) is related to the underlying switch causes.
Unintentional
5
27
32
22
Usual engine
4
13
33
20
As mentioned earlier, the causes were collected via the SwitchWatch questionnaires presented to participants when they were
about to switch search engines. The candidate set of causes was
derived from previous work and included: (i) dissatisfaction with
current engine, (ii) additional topic coverage or verification, (iii)
unintentional, (iv) post-switch engine was better for this type of
task, (v) returning to usual engine, and (vi) other.
Better for task type
12
8
12
11
In Figure 3 we present the six causes of search engine switching
for each class of query changes (SQ, RQ, DQ), and across all
switches (All). The colors correspond to those used in Figure 1,
but unlike Figure 1, these explanations were captured at switch
time rather than based on users’ recollection of switching events.
Focusing on the distribution for All in Figure 3, we compare the
in-situ reasons and the retrospective reasons reported in Figure 1.
There are differences in the figures. For example, dissatisfaction
and coverage were less common in-situ than retrospective, and
reasons associated with user preferences and unintentional or
other reasons were more common in-situ than retrospective. There
are some differences in the questions used in the two cases, which
make it challenging to map directly between them, as well as a
different set of respondents. However, there are some interesting
relations, e.g., dissatisfaction in Figure 1 includes “dissatisfied”
(24%), “expected better” (23%), and “frustrated” (10%), but in
Figure 3, dissatisfaction is only “dissatisfied” (26%) and corresponds well with the 24% from the retrospective study. Alternatively, the SQ results in Figure 3 correspond well with results in
Figure 1. One explanation could be that when people consider
switching retrospectively, they focus on SQ instances.
There are also different distributions of reasons for switching for
the different query relations. The reasons for SQ switches are
primarily dissatisfaction and coverage, whereas preferences (usual
engine and better for task type) and other (unintentional and
other) are the primary reasons for DQ. The reasons for RQ lie
between these two extremes. Interestingly, as shown in the figure,
most of the DQ switches are unintentional or indicate a return to
respondents’ preferred search engine, and only 5% of them were
caused by dissatisfaction. This suggests that DQ switches may be
misleading if used in deriving performance-related metrics,
although they are interesting to understand general preferences.
Figure 3. Relationship between query changes and switch
causes. Values are shown in the bars in the same order as
in the table below the chart.
In contrast, SQ has the greatest fraction of queries associated with
intentional causes such as dissatisfaction and coverage. Switches
caused by dissatisfaction are perhaps the most interesting to a
search provider; pre-switch queries are those that the origin
engine can improve on in order to retain users, while post-switch
queries are important for the destination engine to perform well on
in order to gain users. Therefore, we focus on SQ switches for
later analysis in predicting switching rationales.
In this section we have studied the change of switching queries,
search goals, and their connections with the causes of engine
switching. We now turn our attention to interaction behaviors
associated with different switching motivations. Knowledge of
these behaviors may help us predict the reasons behind engine
switches given observed session-level search interactions.
5.4 User Behaviors for Same-Query Switches
We study the behaviors surrounding engine switching using the
following three groups of features: (i) query, (ii) pre-switch behavior, and (iii) post-switch behavior. These features are useful in
characterizing switching and may be useful for predicting switching rationales. All of the features used are described in Table 3,
along with the mean and the standard deviation feature values
from the in-situ log data for switches associated with dissatisfaction (DSAT), Coverage, and all other reasons (Other). For each
feature we perform one-way independent measures analyses of
variance (ANOVA) between the three causes of switching, and
indicate in the table which paired differences are significantly
different with Tukey post-hoc testing, for significance with <
.01 and < .05. We now describe each feature group in more
detail, with reference to Table 2 where appropriate.
Query features derived from the query string itself, include the
query length in words and characters, and the time between the
pre- and post-switch queries. The longer the query, the more likely the searcher is to be dissatisfied with the search results. Table 2
shows that the number of words in queries associated with dissatisfaction is approximately four, while number of words for
switching queries for coverage and other are around three, although the results are not reliable statistically. This is consistent
with previous work showing that query length can be an important
determinant of search success in Web search [1]. Also included is
the time between the pre-switch and post-switch query. The time
between the queries is longer for switches associated with dissatisfaction, perhaps because the user is considering all top-ranked
search results before making the decision to switch engines.
Pre-switch behavior features from the search interaction before
the engine switch include the number of queries (total, unique,
and reformulations), the number and rate of clicks broken down
by satisfaction and bounce, and the length of the post SERP-trail.
Satisfaction with a clicked result (denoted Sat in Table 2) is defined to be a dwell for longer than 30 seconds, as in previous work
in this area by Fox et al. [9]. Bounces are defined as clicks with a
dwell time on the landing page of under 15 seconds, where the
searcher returns to the SERP after viewing the landing page. The
number of queries is higher for switches associated with coverage,
significantly so for the number of queries (pre_q). In coverage
scenarios, users visit more search results (pre_c) and tend to dwell
longer on these results (pre_c_SAT). We would also expect that
the smaller the number of the satisfied clicks and the more bounces, the more likely the user is dissatisfied. Trends are in this direction, significantly between the number of bounce clicks
(pre_c_Bounce) for DSAT switches and Other. Reformulation
rates (pre_reformRate) are also somewhat higher for DSAT versus
Other. Also of interest are the number of pages on trails following
the SERP click and the number of pages on those trails with the
query in their title, the latter of which (pre_c_containsQ) appears
slightly higher for coverage switches, although not significantly.
Post-switch behavior features correspond to the pre-switch behaviors described earlier in this section. Post-switch behaviors can
provide insight into the nature of the search task, and hence potentially the reason for the search engine switch. Interestingly, none
of the differences for post-switch behavior were statistically significant. This suggests that post-switch behavior may be less useful for differentiating between switching motivations, something
that we will return to later in the paper when we discuss prediction. That said, trends in the findings revealed some similarities to
pre-switch behaviors. For example, in switches associated with
coverage, there are more queries in the session (post_q), and more
query reformulation for DSAT switches (post_reform).
There are noticeable (and some significant) differences in the
interaction behavior associated with different switching rationales.
In addition to characterizing the interactions associated with the
different rationales, we were also interested in whether we could
predict the reasons behind switching given evidence of searcher
interaction behavior within a search session. We now describe our
work on predicting engine switching rationales using interactions.
6. PREDICTING SWITCHING CAUSES
We built and evaluated classifiers to predict the reasons for search
engine switching. For each reason, we formulate the prediction
task as binary classification, where the goal is to predict whether
an observed switch is attributable to the reason of interest. We
also experimented with multi-class (tertiary) prediction, where the
goal was to correctly attribute one of three switching explanations
—DSAT, Coverage, and Other (everything else)—to an observed
switching event. In this section we describe the results from our
Table 2. Features of search behavior per switching cause.
Bolded features exhibit statistically-significant differences.
Dissatisfaction (DSAT) versus Other: n <.05; o <.01;
Coverage versus Other:  <.05; <.01.
Feature
Mean (stdDev)
DSAT Coverage Other
Query features
q_charLength:
Num. chars in switching query
q_wordLength:
Num. words in switching query
q_timeDiff: Time in seconds between
the pre-switch and post-switch queries
27.3
(±22.5)
4.2
(±3.2)
80.9
(±134)
18
(±13.5)
3.2
(±2.4)
52.5
(±98.9)
18.2
(±12.5)
2.9
(±2.0)
46.7
(±80.1)
2.7
(±2.0)
2.0
(±1.5)
1.0
(±1.5)
0.8
(±0.3)
0.2
(±0.3)
2.0
(±2.9)
0.8
(±1.7)
0.8
(±1.2)
0.9
(±2.1)
0.2
(±0.4)
0.3
(±0.4)
1.4
(±2.1)
4.0
(±3.5)
2.1
(±1.9)
1.2
(±1.9)
0.7
(±0.3)
0.2
(±0.3)
2.0
(±2.5)
1.0
(±1.7)
0.6
(±1.0)
0.4
(±0.5)
0.3
(±0.4)
0.2
(±0.3)
1.6
(±2.1)
2.1
(±1.8)
1.4
(±1.0)
0.4
(±1.0)
0.8
(±0.3)
0.1
(±0.2)
0.7
(±1.1)
0.4
(±0.9)
0.2
(±0.5)
0.3
(±0.4)
0.2
(±0.4)
0.2
(±0.4)
0.5
(±0.9)
0.6
(±2.9)
1.7
(±4.4)
0.4
(±1.8)
2.5
(±1.9)
1.8
(±1.3)
0.8
(±1.3)
0.8
(±0.3)
0.2
(±0.3)
2.3
(±2.4)
0.9
(±1.2)
1.0
(±1.2)
1.0
(±0.4)
0.3
(±0.4)
0.3
(±0.4)
2.0
(±2.3)
2.9
(±3.7)
1.4
(±1.1)
0.5
(±1.1)
0.7
(±0.3)
0.1
(±0.2)
2.1
(±2.8)
0.8
(±1.2)
1.0
(±1.6)
0.8
(±0.5)
0.3
(±0.3)
0.3
(±0.4)
1.8
(±2.6)
2.2
(±2.5)
1.3
(±0.8)
0.4
(±0.8)
0.8
(±0.3)
0.1
(±0.2)
1.6
(±2.0)
0.7
(±0.9)
0.7
(±1.0)
0.9
(±0.5)
0.3
(±0.4)
0.3
(±0.4)
1.2
(±1.8)
1.6
(±3.9)
1.3
(±5.9)
3.5
(±8.9)
Pre-switch features
pre_q: Num. queries in session
pre_uniqQ:
Num. unique queries in session
pre_reform: Num. query reformulations
pre_uniqQRate: pre_uniqQ / pre_q
pre_reformRateo : pre_reform / pre_q
pre_c: Num. SERP clicks for
related queries
pre_c_Sat: Num. satisfied SERP clicks
for related queries
pre_c_Bounceo : Num. bounce SERP
clicks for related queries
pre_cRate: pre_c / pre_q
pre_c_SatRate: pre_c_Sat / pre_q
pre_c_BounceRate:
pre_c_Bounce / pre_q
pre_t: Num. pages on click trail
pre_c_containsQ: Num. SERP clicks
on a search result with title containing
at least one non-stop-word query term
Post-switch features
post_q: Num. queries in session
post_uniqQ: Num. unique queries
post_reform: Num. query reformulations
post_uniqQRate: post_uniqQ / post_q
post_reformQRate:
post_reform / post_q
post_c:
Num. SERP clicks for related queries
post_c_Sat: Num. satisfied SERP clicks
for related queries
post_c_Bounce: Num. bounce SERP
clicks for related queries
post_cRate: post_c / post_q
post_c_SatRate: post_c_Sat / post_q
post_c_BounceRate:
post_c_Bounce / post_q
post_t: Num. pages on click trail
post_c_containsQ: Num. SERP clicks
on a search result with title containing
at least one non-stop-word query term
experiments. We begin by describing the classification algorithm
used in the prediction, then describe the evaluation metrics, the
models compared in the study, and then the prediction findings.
models via ten-fold cross validation, across 100 randomized
experimental runs, and report averages across all runs and folds.
6.1 Classifiers
We begin our analysis by comparing the performance of each of
the binary classification algorithms with the two baselines, for
each of the three classes: DSAT, Coverage, and Other. Table 3
reports the average
values for each switching explanation
versus baselines for all features listed in Table 2, and the results of
paired -tests between the models and the baselines.
We used features described in Table 2 and experimented with a
variety of different classification algorithms for predicting switching causes, including decision trees, logistic regression, and naïve
Bayes [5], which were the three best performers. The performance
of all three methods was similar, and we report on the results of
the logistic regression classification here.
6.2 Evaluation Metrics
In evaluating the performance of our predictions, we measure
precision (the fraction of predicted instances that were correctly
predicted) and recall (the fraction of all true instances that were
correctly predicted). We report on the
measure, with set to
0.5, which gives twice as much weight to precision than to recall.
Precision is very important in application scenarios for a predictor
of switching rationales. In an online scenario, we would want to
be highly confident before adapting the search experience based
on switch rationale predictions. In an offline scenario, such as
studying dissatisfied switches in log data, we need to obtain a set
of dissatisfied switches for further analysis. Since there are many
switching events in logs, we do not need to classify all switches
(have high recall) as long as we can precisely label some.
6.3 Methods Compared
We compare a number of different methods for predicting switching rationales. We used two strong baselines which leverage the
marginal distribution and use rules derived manually from a visual
inspection of switching events in the logs. The baselines are:
• Baseline (Prior): Bases predictions on the class distribution.
• Baseline (Rule): Uses rules derived from inspection of switching events in logs. Predict DSAT if there is no click before the
switch and one or more clicks after the switch; predict Coverage if there are clicks before and after the switch; predict Other
(i.e., all reasons other than dissatisfaction or coverage) if neither of the above rules are triggered.
In addition to these baselines, we also trained binary and tertiary
classifiers on varying sets of features described in Table 2:
• All: Classifiers trained on all features.
• Query: Classifiers trained only using query features.
• Pre-switch: Classifiers are trained only using pre-switch behavioral features. These are the features available before engine
switching and could be used in combination with a switch predictor (such as that described in [22]) to predict the reason for
an anticipated switch.
• Post-switch: Classifiers trained only using post-switch behavioral features. This could help the destination engine predict the
reason for the incoming switch and adjust the search experience
accordingly (e.g., provide diverse results if reason is coverage).
We now present findings on prediction effectiveness using the
different feature classes. Given the importance of SQ switches to
search providers (as discussed earlier in Section 5.3.3), we elected
to focus on the set of 354 in-situ SQ switches, and used them for
training and testing. This is a relatively small set given the fairly
intensive SwitchWatch deployment effort, primarily because
engine switching is a rare event and to maintain reaslism, we did
not artificially promote switching in our study. We compare the
6.4 Binary and Tertiary Predictions
Table 3. Binary prediction performance (measured via
) of
DSAT, Coverage, and Other. Significance of differences between models and baselines is marked: Baseline (Prior):
<.05, <.01; Baseline (Rule):  <.05, <.01.
Method
DSAT
Coverage
Other
Base (Prior)
72.40
27.12
17.40
Base (Rule)
48.84
24.19
20.20
All Features
85.69
47.84
29.01
The findings presented in Table 3 above show that the prediction
model trained with all features significantly outperforms both
baselines in the prediction of DSAT and Coverage, and marginally
outperforms the baselines in the prediction of Other. The observed
gains over the baselines are strong given the limited amount of
data available and suggest that there is good predictive signal in
the features, espcially for DSAT predicitons which are of great
interest to search engine providers. Prediction of the Other class
appears more challenging, perhaps because this class includes
several different switching motivations, each of which may have
its own associated behavioral patterns.
In addition to the binary classification, we also experimented with
multi-class prediction of switch explanation among three reasons:
DSAT, Coverage, and Other. Multi-class prediction is important
because it allows search providers to use a single predictor of
switching reasons, potentially reducing training and deployment
overhead. Findings from prediction experiments conducted in the
same way as above (but this time with three-level judgments)
show that the
of a logistic regression model (74.58) exceeded
both of the baselines (Prior: 59.46, Rule: 52.80) at < .01. This
suggests that multi-class prediction using in-situ data is feasible
with these data. However, further work is needed to establish
whether a multi-class predictor would outperform a combination
of binary classifiers, such as those described above, and study the
cost-benefit tradeoffs of each solution.
6.5 Feature Group / Feature Performance
In Table 4 we present the average
metric for how well each
model predicts DSAT for different sets of features: all features,
pre-switch features, query features, and post-switch features.
Table 4. Feature group performance (measured via
for in-situ assessment (DSAT).
)
Group
All
Pre-switch
Query
Post-switch
F0.5
85.69
81.12
74.28
78.99
All differences between the four feature groups and the baselines
were statistically significant using paired -tests at < .05. The
best predictive performance was attained when all features were
used. The most important features learned for those predictions (in
descending order of importance) and directionality of their DSAT
relationship were pre_c_SatRate (−), pre_c (+), pre_uniqQRate
(+), post_c_containsQ (+), and post_reformRate (−). SERP clicks
and query reformulation behavior on both the pre- and post-switch
engines may be predictive of user satisfaction, a claim supported
by [11]. Factors such as low satisfaction on the pre-switch engine
were also good predictors of dissatisfaction-related switching.
Pre-switch features appeared to provide more predictive signal
than post-switch features, as was suggested in our earlier analysis
of those search behaviors (see Section 5.4). Pre-switch interaction
behavior reveals more about searchers’ behavior leading up to the
switching event, and therefore might provide better quality evidence of the reason behind the switch. In addition, it may also be
that there is more variance in what users do following a search
engine switch, and that makes features of post-switch behavior
less reliable indicators of engine switching rationale.
In-situ assessment affords the capture of switching explanations at
switch time from the searcher, and should be an accurate elicitation method for switching rationales. However, it may not always
be desirable to deploy such a tool to searchers, especially on a
large scale, given interruption costs to users, privacy implications,
and the infrastructure required to store large volumes of behavioral data. Therefore, we also explored the use of editorial assessments by third-party judges performing manual log analysis to
identify the reasons for search engine switching. Including editorial assessments in our study helps us better understand the judgment correspondence between judges and switching users.
7. EDITORIAL ASSESSMENTS
The main advantages of editorial assessments lie in the large volume of switching data available for analysis and the fact that labels can be obtained faster than the in-situ method. In addition,
sessions drawn from logs may be more representative of general
Web search activity than those from a subset of searchers who
elect to participate in in-situ switch monitoring. The disadvantages of editorial assessment include the possible misinterpretation of switching rationales by third-party human assessors and
the cost of human labor involved in performing the assessments.
Therefore, in this section we consider the relation between in situ
judgments by the searcher and those by third-party assessors.
7.1 Log Data
We randomly selected 100 search sessions containing at least one
same-query engine switch from the in-situ log data used so far in
the paper. Each of these switches was judged by two human assessors and a switching reason was assigned to the first samequery switch in the session. We selected sessions from the in-situ
logs so we could directly compare the two judgment methods on
the same set of switches. As we will describe later in the section,
the judgment task was very intensive. To maintain judgment quality we focused on same-query switches to allow judges to focus
on identifying the switching reason (rather than goal changes,
etc.), restricted judgments to the first same-query switch in each
session, and limited the number of judgments per judge to 100.
7.2 Judges
Two human judges performed the editorial assessment task. Both
judges are Web search researchers who are familiar with engine
switching and search log data. Initial training and discussion was
conducted as a pair to help ensure consistency in the labeling. The
group examined a few example sessions containing switching
events and discussed likely switching motivations.
7.3 Editorial Guidelines
Judges were presented with a spreadsheet containing the 100 sessions to be judged. Each row contained a unique session identifier
and the URL of the page visited. If the URL was a search engine
result page, the query and engine name were also shown, as were
timestamps, and browser and tab identifiers, to help track multitasking. Judges answered a few intermediate questions about the
search sessions before identifying the reasons for engine switching in order to get acquainted with the searcher’s intent and overall experience with the search engine. The intermediate questions
included: determining the search goal of the user (based on the
taxonomy of search goals proposed by Rose and Levinson [19]),
whether the information need requires multiple sources to fulfill,
overall success, pre-switch success, and post-switch success. To
answer these questions, judges used the information about landing
page content and how well it matched query terms as well as the
actual sequence of query and click patterns on search results in a
session. When determining the reason for the engine switch, the
response options included: Dissatisfaction, Coverage, and Other.
Note that we regard unintentional, user habit or user preferences
(highlighted as being an important reason in Figure 1) as a subset
of Other here, since judges may not be able to assess them from
log data based solely on a single-session evidence.
Answer options for all questions were presented to judges as dropdown lists in the judgment spreadsheet. Space was also provided
for additional comments, although this was seldom used in practice. Judges performed their judgments in isolation and then met
to discuss and resolve inconsistencies.
7.4 Judge Agreement
Each judge assessed the same 100 search sessions and each session contained at least one SQ switch (that first of which was labeled). The Cohen’s kappa ( ) between the two judges for judging
switching reasons as Dissatisfaction, Coverage, and Other was
0.78 while between judges for Dissatisfaction and Other, where
we merged Coverage and Other classes, was approximately 0.88.
This signifies “substantial” judge agreement for tertiary labeling
and “almost perfect” judge agreement for binary labeling [14].
7.5 Switch Causes for Same-Goal Switches
In the case of editorial assessments, we regarded switches with the
same query as having the same search goal (given that the analysis in the previous section showed that SQ was a reasonable proxy
for SG). As mentioned earlier, in this analysis we only focus on
Dissatisfaction, Coverage, and Other (i.e., all reasons that are not
dissatisfaction or coverage).
Overall, there was an 83% agreement between the reasons provided by the third-party judges and the switchers over the same 100
switches used in the analysis presented in this section. The typical
disagreements lay in Coverage and Other, with it being most challenging to differentiate coverage and preference-based switches,
which were present in Other but not explicitly labeled by judges.
There appears to be reasonable agreement between in-situ and
editorial sources on the three main reasons for switching for the
SQ switches. The editorially-assessed data has a similar distribution to the in-situ data: the percentage of switches associated with
dissatisfaction is similar but slightly lower (45% vs. 49%) and the
percentage of switches associated with coverage is similar but
higher (33% vs. 21%). The differences in coverage estimates may
be real, or given fewer assignable switching reasons (three as
opposed to six in the in-situ survey), our judges may have overestimated the amount of switching associated with topic coverage.
7.6 Predicting Switching Causes
Prediction models were constructed using the data provided from
the editorial assessment process. In Sections 7.6.1 and 7.6.2 we
train and test on editorial judgments. However, in Section 7.6.3,
we train on editorial judgments and test on in-situ judgments.
7.6.1 Binary and Tertiary Predictions
Baselines were updated to reflect the distributions in the editorial
data set. Table 5 shows the average obtained
values. The
performance of the prediction model based on all interaction features is reasonable, outperforming both baselines for DSAT and
Coverage, and performing marginally better for Other. Performance on predicting DSAT exceeds that of predicting coverage,
but the observed gains over the baselines in both cases are lower
than we observed for in-situ assessments. One reason for this
difference is that less data were available for training and testing
the predictive models (i.e., 100 editorial judgments vs. 354 in-situ
judgments). To test the extent of this effect, we trained and tested
predictive models on the same 100 in-situ switches (using ten-fold
cross validation and 100 runs, as before), and observed small differences: the DSAT
value with 100 in-situ judgments was
79.43 vs. 85.69 with 354 in-situ judgments. Another reason for the
lower performance with editorial assessment could be noisy labels
from third-party judges, who may incorrectly interpret logs and
would make it challenging to associate reasons with actions.
Table 5. Binary prediction performance (measured via
) of
DSAT, Coverage, and Other. Train on editorial, test on
editorial. Symbol meaning same as Table 3.
Method
DSAT
Coverage
Other
Baseline (Prior)
50.41
35.54
30.16
Baseline (Rule)
54.88
41.90
30.98
All Features
66.16
47.44
33.86
In addition to the binary classification, we again experimented
with multi-class predictions, this time with the editorial data. The
findings showed significant gains in
over baselines (Logistic
regression: 64.40, Prior: 50.10, Rule: 45.45) at < .01, suggesting
that tertiary prediction using editorial data may also be feasible.
7.6.2 Performance of Feature Groups and Features
In a similar way to Section 6.5, we now examine performance on
predicting DSAT this time using editorial assessment data. Table 6
summarizes performance for all features, for only query features,
and for only pre- and post-switch interaction features.
Table 6. Feature group performance (measured via
for editorial assessment (DSAT).
)
Group
All
Pre-switch
Query
Post-switch
F0.5
66.16
64.50
57.69
60.12
Once again, prediction performance is worst when using queryonly and post-switch features, suggesting that they may be least
useful for reliably predicting switch causes. We have already noted that post-switch behavior can be highly variable, making predictions based on it challenging. In addition, in post-assessment
debriefings we discovered that judges generally ignored postswitch behavior; something that needs to be resolved in future
studies, perhaps by modifying judge instructions.
7.6.3 Predicting In-Situ Judgments
In addition to using the editorial judgments to predict editorial
judgments, we can also use the editorial judgments to predict the
in-situ judgments, under the assumption that the in-situ judgments
are the real ground truth. We re-ran our experiments, again using
ten-fold cross validation over 100 runs, but instead of using the
editorial judgments for the test fold, we used the associated in-situ
judgments. The performance findings are reported in Table 7.
Table 7. Binary prediction performance (measured via
) of
DSAT, Coverage, and Other. Train on editorial, test on in-situ.
Symbol meaning same as Table 3.
Method
DSAT
Coverage
32.00
Other
Baseline (Prior)
47.83
25.05
Baseline (Rule)
50.65
37.17
27.71
All Features
62.76
44.33
31.92
The findings show that the performance of the editorial judgments
in predicting the in-situ judgments is slightly lower than that obtained when predicting editorial judgments. One explanation is
differences in the criteria used by the judge and by the switcher;
such inconsistency would lead to poor predictive performance on
this task. More work is needed to understand these and other differences noted in this section, primarily because third-party labeling of switching episodes would likely be used in practice.
8. DISCUSSION AND IMPLICATIONS
We have presented an investigation into the causes of search engine switching and the automatic prediction of switching rationales as identified via in-situ and editorial assessments. The study
has provided valuable insights into the reasons behind search
engine switching and shown that we can predict the motivation
behind engine switching with only limited interaction evidence.
We found that a large percentage of DQ switches are unintentional or preference-related. Therefore, it could be misleading if these
switches were included in performance-related analysis and metric derivation. As expected, search goals typically remain constant
during SQ search engine switches, there are some behavioral patterns (especially in pre-switch behavior) that can reveal different
motivations for SQ switching, and affording the accurate prediction of different switch rationales. Our findings provide better
understanding of switching, and help search engines improve their
user experience or derive more accurate competitive metrics.
The findings of the prediction experiments revealed that using the
behaviors both before and after the switch lead to the most accurate predictive performance, with accuracy ranging from 65-85%
depending on the source of the judgment data. The analysis also
showed that the most predictive subset of features were those
from interaction preceding the switch. This is promising for the
development of real-time support for dissatisfied users, and is in
line with the findings of Feild et al. [8], who demonstrated value
in using recent behavior (including those associated with switching) to predict frustration. In future work, we will experiment
further with query features such as type or aggregate clickthrough
rate, which have been shown to be effective in predicting query
performance and may also be useful in this context [10]. Given
that dissatisfaction with the pre-switch engine was the dominant
switching rationale, we will also explore the reasons that underlie
user dissatisfaction (e.g., lack of diversity, obsolete results, etc.).
The participants in our study were all Microsoft employees. Although we showed some similarities between the switching behav-
ior of these users and those reported in previous work [22], a larger deployment of SwitchWatch beyond our organization to a more
diverse cohort is needed to further generalize our findings.
This work has a number of practical implications for the design of
user-facing search technology. First, predictions of the reasons for
switching can be used to dynamically adapt the search experience.
Previous work has shown that we can accurately predict when a
user will switch engine [12][15]. In this research we have shown
that we can predict the reason behind such switching with good
accuracy. Combining these methods would enable us to predict in
real-time when a user is going to switch engines because they are
dissatisfied. Over time, dissatisfaction-related switches can potentially erode user confidence in the search engine, and ultimately
lead to permanent switching. Advanced warning of when switches
are likely to occur enables search engines to intervene with an
improved experience or offer new capabilities to candidate dissatisfied switchers. Such advanced capabilities include real-time chat
with a domain expert or more powerful (but also more computationally-costly) search technologies. Conversely, in cases where
the switch is detected by the post-switch engine, perhaps through
a toolbar or inspecting the referrer URL, the engine could predict
the reason for the switch to them based on post-switch behavior.
For example, for incoming switchers dissatisfied with the preswitch engine, more computational resources could be devoted to
ranking. For incoming switchers seeking topic coverage or verification, emphasis could be put on search result diversity.
In addition, search engine companies extensively analyze log data
to identify opportunities for improving their service. The ability to
predict the reasons for switching allows search engines to compute more accurate competitive metrics that target for improvement of queries frequently leading to dissatisfied switching.
9. CONCLUSIONS AND FUTURE WORK
We have presented a study of the reasons that people switch
search engines within sessions. We capture the reasons for switching and associated search behaviors in-situ and use the data to
develop and evaluate models to automatically predict switching
motivations using features of the switching queries and pre- and
post-switch behavior. Our findings offer insight into searchers’
decision-making processes and demonstrate the relationship between behaviors and switching causes. The findings also reveal
sufficient consistency in search behaviors to afford accurate prediction of switching reasons. This could be useful for search providers to improve the search experience for users and derive more
accurate competitive metrics. Future work involves studying the
different types of dissatisfaction-related switching in more detail,
exploring the use of additional features for predictions, creating
refined metrics, and deploying switch rationale predictors on a
search engine to tailor search experiences to switch rationales.
REFERENCES
[1] Aula, A., Khan, R. M., and Guan, Z. (2010). How does
search behavior change as search becomes more difficult?
Proc. CHI, 35-44.
[2] Bailey, P., White, R.W., Liu, H., and Kumaran, G. (2010).
Mining past query trails to label long and rare search engine
queries. ACM TWEB: 4(15).
[3] Bartell, B.T., Cottrell, G.W., and Belew, R.K. (1994). Automatic combination of multiple ranked retrieval systems.
Proc. SIGIR, 173−181.
[4] Belkin, N., Cool, C., Croft, W.B., and Callan, J. (1993). The
effect of multiple query representations on information retrieval system performance. Proc. SIGIR, 339−346.
[5] Bishop, C.M. (2006). Pattern Recognition and Machine
Learning. New York: Springer.
[6] Carmel, D., Yom-Tov, E., Darlow, A., and Pelleg, D. (2006).
What makes a query difficult? Proc. SIGIR, 390−397.
[7] Cronen-Townsend, S., Zhou, Y. and Croft, W. B. (2002).
Predicting query performance. Proc. SIGIR, 299−306.
[8] Feild, H., Allan, J., and Jones, R. (2010). Predicting searcher
frustration. Proc. SIGIR, 34−41.
[9] Fox, S., Karnawat, K., Mydland, M., Dumais, S.T., and
White, T. (2005). Evaluating implicit measures to improve
the search experience. ACM TOIS, 23(2): 147−168.
[10] Guo, Q., White, R.W., Dumais, S.T., Wang, J., and Anderson, B. (2010). Predicting query performance using query,
result, and user interaction features. In Proc. RIAO.
[11] Hassan, A., Jones, R., and Klinkner, K.L. (2010). Beyond
DCG: User behavior as a predictor of a successful search.
Proc. WSDM, 221−230.
[12] Heath, A.P. and White, R.W. (2008). Defection detection:
Predicting search engine switching. Proc. WWW,
1173−1174.
[13] Juan, Y.F. and Chang, C.C. (2005). An analysis of search
engine switching behavior using click streams. Proc. WWW,
1050−1051.
[14] Landis, J.R. and Koch, G.G. (1977). The measurement of
observer agreement for categorical data. Biometrics, 33,
159−174.
[15] Laxman, S., Tankasali, V., and White, R.W. (2008). Stream
prediction using a generative model based on frequent episodes in event sequences. Proc. SIGKDD, 453−461.
[16] Leskovec, J., Dumais, S., and Horvitz, E. (2007). Web projections: Learning from contextual subgraphs of the Web.
Proc. WWW, 471−480.
[17] Mukhopadhyay, T., Rajan, U., and Telang, R. (2004). Competition between internet search engines. Proc. HICSS.
[18] Pew Internet and American Life Project. (2005). Search Engine Users. Accessed December 15, 2008.
[19] Rose, D. and Levinson, R. (2004). Understanding user goals
in Web search. Proc. WWW, 13−19.
[20] Teevan, J., Dumais, S., and Liebling, D. (2008). To personalize or not to personalize: Modeling queries with variation in
user intent. Proc. SIGIR, 620−627.
[21] Telang, R., Mukhopadhyay, T., and Wilcox, R. (1999). An
empirical analysis of the antecedents of internet search engine choice. Proc. Wkshp on Info. Systems and Economics.
[22] White, R.W. and Dumais, S.T. (2009). Characterizing and
predicting search engine switching behavior. Proc. CIKM,
87−96.
[23] White, R.W., Kapoor, A., and Dumais, S.T. (2010). Modeling long-term search engine usage. Proc. UMAP, 28−39.
[24] White, R.W., Richardson, M., Bilenko, M., and Heath, A.P.
(2008). Enhancing Web search by promoting multiple search
engine use. Proc. SIGIR, 43−50.
[25] Zhou, Y. and Croft, W.B. (2007). Query performance prediction in Web search environments. Proc. SIGIR, 543−550.